Audited Conformal Prediction for Classification under Unknown Distribution Shift
Researchers have introduced Audited Conformal Prediction (ACP), a novel method designed to improve uncertainty quantification for classification models facing unknown distribution shifts. ACP utilizes a small target dataset to train an auxiliary model that identifies potential failures of the pre-trained model. By integrating this audit model into the conformal prediction framework, ACP aims to provide prediction sets with guaranteed marginal coverage and enhanced conditional coverage. AI
IMPACT Enhances reliability of deployed classification models by improving uncertainty quantification under distribution shift.